Cheat sheet

Tools

From lesson 0005. The model writes a slip. You cook. The ticket re-enters RAM.

1. ProposeStructured tool call, not English.
2. RunYour code or an MCP server.
3. PasteResult goes into the script.
4. PickPack · read · pick, now with a live fact.

When

• Do: live state, actions, anything the weights cannot know tonight.
• Skip: already in the script, or a tiny general question (2+2).
• RAG pastes pages you found. Tools paste tickets from work you ran.

MCP

USB-C for AI apps: one plug, many kitchens. Same loop. Not a bigger window, not a new tokenizer.

OpenAI: tool call → you execute → tool output. Anthropic: the model never executes on its own.

The loop in code

From lesson 0021. Claude Messages API shapes; TypeScript.

const messages = [{ role: "user", content: userText }];
let turns = 0;

while (true) {
  if (turns === maxTurns) return "max_turns";            // CAP
  turns++;

  const response = await client.messages.create({ model, max_tokens, tools, messages });

  if (response.stop_reason === "end_turn") return text;   // STOP

  messages.push({ role: "assistant", content: response.content });   // KEEP, whole

  const results = response.content                         // RUN
    .filter((block) => block.type === "tool_use")
    .map((slip) => ({
      type: "tool_result",
      tool_use_id: slip.id,
      content: runTool(slip.name, slip.input),
    }));

  messages.push({ role: "user", content: results });       // PASTE, one message
}

• stop_reason: tool_use means run slips; end_turn means done.
• Results go in a user message, immediately after the slip message. No tool role.
• Every tool_use id needs a matching tool_result, or the API returns 400.
• The whole array is resent each call. Input tokens grow every turn.

Source: Handle tool calls.